Work

On Prognosis Modeling

Year: 2022
Type: dissertation
Abstract: With the rising adoption of Eletronic Health Record (EHR) systems in healthcare facilities, we are presented with an unprecedented amount of rich information regarding patient care and the evolution of health conditions. This has given rise to research opportunities to model hospital data for several patient trajectory tasks, specifically predicting future diseases. Prognosis modeling is the task of predicting future diagnoses a patient will be as- signed to in the future, given their medical history. If a model can learn the temporal development of diagnosed health conditions, then hospitals are better prepared to pro- vide early recognition and treatment of patients. In this dissertation we explore different deep learning and machine learning models for the task of prognosis modeling. We make a thorough analysis and discussion on how to build a prognosis dataset from an EHR dataset, and work with two datasets: a publicly available EHR dataset Medical Information Mart for Intensive Care (MIMIC-III) of around 50 thousand patients, and a private dataset from Hospital da Luz (HDL) with almost half a million patients. On both datasets we conclude that both the sequence model (RNN) and the basic Multi-Layered Perceptron (MLP) model can’t achieve higher performance than the Logistic Regression model, and that all three models mostly learn the identity function for chronic conditions as well as some weak co-occurrence functions, while having very poor performance on predicting onset of new conditions for the patient. We also concluded our models were well-calibrated, and leveraged this finding to both provide a baseline abstention pipeline of uncertain predictions, as well as make the case for the importance of well-calibrated models in healthcare. Lastly, we developed Monte Carlo Dropout models to make use of the posterior predictive distribution as a proxy to uncertainty of predictions, and did not achieve significant improvements.
Language: English
Open Access status: green
OpenAlex ID: W7140349251
Cites: 52
Cited by:
Related to: 10